JP2012022530A - Behavior prediction apparatus, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To properly predict a future position of a user even when he or she behaves in an unusual way, and predicts the time in addition to the position.SOLUTION: A behavior prediction apparatus comprises the steps of: generating state transition system data sampled at predetermined time intervals on the basis of a past position data history acquired from a mobile terminal of a user; generating event system data on the basis of scheduler data acquired from the mobile terminal; embedding the state transition system data, event system data, time zone data and stay time length in each state, as probability variables, into a DBN model to estimate a parameter that defines probability distribution in a probability model; and adding the current position of the user, the future scheduler data and the estimated parameter to the DBN model to determine and output a state transition system for the maximum probability as information indicating a result of predicting user behavior.

Description

  The present invention relates to an action prediction apparatus, method, and program for predicting a user's future action based on the user's position information and schedule information.

  A method for predicting user behavior by constructing a prediction model based on acquired and accumulated position data by carrying a mobile terminal equipped with a position measurement device typified by GPS (Global Positioning System). Has been proposed. For example, in Non-Patent Document 1, the location where the user stayed is extracted from the acquired GPS data by clustering, and a Markov model representing movement between the stay locations is constructed based on the accumulated data to predict the behavior. The technique to do is described.

  Similarly, in Non-Patent Document 2, after extracting the place where the user stayed by clustering GPS data in the same manner, the daily behavior of the user is expressed as a series in which the stay places are arranged, and then this series is expressed. A method is described in which a sequence pattern is extracted by applying a sequence pattern mining process, and a user's behavior is predicted by using the extracted pattern.

  If the user's behavior can be accurately predicted, more useful information can be provided to the user. For example, by predicting the return route when the user is at the company, information that reads ahead one step, such as providing disaster information on the predicted route and time sale information of stores that exist along the route in advance Provision is possible.

Daniel Ashbrook and Thad Starner, “Using GPS to Learn Significant Locations and Predict Movement across Multiple Users”, Personal and Ubiquitous Computing 7 (5), pp. 275-286 (2003) M. Nishino, T. Yamada, S. Seko, M. Motegi, S. Muto and M. Abe, A place prediction algorithm based on frequent time-sensitive patterns, In Pervasive 2009 adjunctive proceedings, 2009

  However, in the conventional method of predicting behavior only from the history of position data, since prediction is performed using only regularity of behavior, there is a limit to prediction accuracy. On the other hand, when the user registers a future scheduled action or the like in the scheduler, the scheduled action entered in the scheduler is highly likely to be executed, and can be accurately predicted. However, when the scheduled action described in the scheduler is not executed as scheduled, or when multiple scheduled actions are registered in the same time zone, the user's action is accurately predicted only from the scheduler. It ’s difficult. In addition, since not all scheduled actions are generally entered in the scheduler, it is difficult to predict continuously what actions the user will take from only the scheduler.

  The present invention has been made paying attention to the above circumstances, and the purpose of this invention is to make it possible to accurately predict the future stay position even when the user performs a different behavior, and not only the position. An object of the present invention is to provide an action prediction apparatus, method, and program which can predict stay time.

  In order to achieve the above object, a first aspect of the present invention is to acquire position data of a mobile terminal possessed by a user in a past first period, and based on the acquired position data, a fixed time interval. Means or process for generating and storing state transition series data represented as a set of the sampling time set in step 1 and information representing the staying state of the user at the time, and the scheduled event of the user in the first period The scheduler data represented by the contents, the scheduled start time and the scheduled end time is acquired, and based on the acquired scheduler data, as a set of the set sampling time and a keyword representing the contents of the scheduled event at the time Means or process for generating and storing represented event sequence data, and the stored state transition sequence data and To event-series data, and means or process for setting a plurality of time zones based on the sampling time included in the respective data, and a means or process performs parameter learning process. Then, in the means or process for performing the parameter learning process, a probability model that represents a probabilistic coexistence relationship between events in time series for the user's behavior prediction is defined, and the stored state transition sequence data and Each event series data is read out, and the read state transition series data, event series data, and the set time zone are incorporated into the probability model as random variables, and a probability distribution is defined in the probability model. The parameters are estimated.

  Therefore, according to the first aspect of the present invention, a probability model that represents a probabilistic coexistence relationship between events in a time series is used for predicting a user's behavior, and the user's location data is included in the probability model. The parameters that determine the probability distribution are estimated by incorporating the state transition sequence data generated based on the event sequence data generated based on the user's scheduler data and information indicating the time zone as random variables. . Therefore, by using this estimated parameter, the user's future behavior, that is, at which time and at which location, considering the history of location data and future scheduled events and time zones entered in the scheduler. It is possible to predict where to stay or from which place to move.

In order to achieve the above object, according to a second aspect of the present invention, in addition to each means or process described in the first aspect, the user stays in each state included in the stored state transition sequence data. A means or process for obtaining a time length is further provided, and in the means or process for performing the parameter learning process, a probability model representing a probabilistic coexistence relationship between events in time series is defined for predicting the user's behavior, The stored state transition series data and event series data are read out respectively, and the read out state transition series data, event series data, the set time zone, and the obtained residence time length are respectively probabilities. As a variable, it is incorporated into the probability model, and parameters that determine the probability distribution in the probability model are estimated and stored.
Therefore, according to the second aspect of the present invention, in addition to the history of the position data, the future scheduled event entered in the scheduler and its time zone, in addition to considering the residence time length in each state, the user's Future behavior can be predicted with higher accuracy.

  When the parameter learning process is executed in the first and second aspects, a parameter that determines the probability distribution may be estimated by a maximum likelihood estimation method.

  The first aspect is represented by a set of the user's action start state and the time based on the acquired current position data of the mobile terminal possessed by the user and based on the acquired current position data. A means for generating start state information, and future scheduler data representing the scheduled event after the user's action start time by its content, scheduled start time and scheduled end time are acquired, and the acquired future scheduler data Based on the means for generating event data represented as a set of a keyword representing the content of the scheduled event and its scheduled execution time zone, and the generated start state information, event data and execution thereof for the probability model Information indicating the scheduled time zone is given as a random variable, and the stored parameter is given to match the state in the probability model. A means for obtaining a state transition sequence when the probability of performing the maximum operation and a means for outputting the state transition sequence that gives the obtained maximum probability as information representing the user's behavior prediction result are further provided. Also features.

  Furthermore, the second aspect is represented by a set of the user's action start state and the time based on the acquired current position data of the mobile terminal possessed by the user and based on the acquired current position data. A means for generating start state information, and future scheduler data representing the scheduled event after the user's action start time by its content, scheduled start time and scheduled end time are acquired, and the acquired future scheduler data Based on the means for generating event data represented as a set of a keyword representing the content of the scheduled event and its scheduled execution time zone, and for the probability model, the generated start state information, event data, and execution thereof Information representing the scheduled time zone and information representing the length of the residence time are given as random variables, and the stored parameters are given to Means for obtaining a state transition sequence when the probability corresponding to the state and its residence time length is maximum in the model, and a state transition sequence that gives the obtained maximum probability as information representing the user's behavior prediction result And a means for outputting.

  If it does in this way, the process which estimates the user's future action in the same apparatus will be performed in the 1st and 2nd viewpoint using the parameter estimated and memorized beforehand by parameter learning processing. It becomes possible.

  That is, according to the present invention, even when a user performs an unusual behavior, a future staying position can be accurately predicted, and not only the position but also the staying time can be predicted, A program can be provided.

BRIEF DESCRIPTION OF THE DRAWINGS The schematic block diagram of the system provided with the action prediction apparatus concerning one Embodiment of this invention. The block diagram which shows the function structure of the action prediction apparatus provided in the system shown in FIG. The flowchart which shows the process sequence and process content of the learning mode by the action prediction apparatus shown in FIG. The flowchart which shows the procedure and content of the parameter learning process performed in the process of the learning mode shown in FIG. The flowchart which shows the process sequence and process content of the action prediction mode by the action prediction apparatus shown in FIG. The flowchart which shows the procedure and content of the probability maximum transition search process in a 1st DBN model performed in the process of the action prediction mode shown in FIG. The flowchart which shows the procedure and content of the probability maximum transition search process in a 2nd DBN model performed in the process process of the action prediction mode shown in FIG. The figure which shows an example of the position data acquired in the process of the learning mode shown in FIG. The figure for demonstrating the concept of the stay location extraction process performed in the process of the learning mode shown in FIG. The figure which shows an example of the stay action data produced | generated in the process of the learning mode shown in FIG. The figure which shows the concept of the state transition series creation process performed in the process of the learning mode shown in FIG. The figure which shows an example of the state transition series data produced | generated in the process of the learning mode shown in FIG. The figure which shows an example of the scheduler data acquired in the process of the learning mode shown in FIG. The figure which shows an example of the event data produced | generated based on the said scheduler data in the process of the learning mode shown in FIG. The figure which shows an example of the event series data produced | generated based on the said event data in the process of the learning mode shown in FIG. The figure for demonstrating the 1st DBN model used for a user's action prediction. It is a figure which shows an example of the correspondence of a time and time. The figure which shows an example of the correspondence of the kind of state which a user can take, and its identification ID. The figure which shows an example of the correspondence of the kind of event content keyword which appears in event series data, and its identification ID. The figure which shows an example of the correspondence of the time slot | zone containing the event time which appears in event series data, and its identification ID. The figure for demonstrating the production process of the learning data in a 1st DBN model. The figure for demonstrating the 2nd DBN model used for a user's action prediction. The figure which shows an example of the correspondence of the kind of user's staying length, and its identification ID. The figure for demonstrating the creation process of the learning data in a 2nd DBN model. The figure which shows an example of the prediction result obtained by the action prediction process shown in FIG.

Embodiments according to the present invention will be described below with reference to the drawings.
FIG. 1 is a schematic configuration diagram of a system including an action prediction apparatus according to an embodiment of the present invention. In this system, a plurality of mobile terminals MS1 to MSn each possessed by a user can be connected to the behavior prediction apparatus SV via a communication network NW.

  The communication network NW includes an IP (Internet Protocol) network and an access network for accessing the IP network. As the access network, a public communication network, a mobile phone network, a LAN (Local Area Network), a wireless LAN, a CATV (Cable Television) network, or the like is used.

  The mobile terminals MS1 to MSn are composed of mobile personal computers such as mobile phones, PDAs (Personal Digital Assistants), smart phones, netbooks, etc., and have standard voice communication functions, mail transmission / reception functions, browser functions, etc. In addition to the functions, it has a position measurement function and a schedule management function (scheduler).

The position measurement function is realized by a GPS (Global Positioning System) receiver, a GPS measurement control unit, and a GPS data transmission control unit.
The GPS receiver receives GPS signals transmitted from a plurality of GPS satellites (not shown) via an antenna. The GPS measurement control unit activates the GPS receiver at a predetermined measurement cycle of, for example, 5 minutes or more, takes in a GPS signal received by the GPS receiver, and generates position data of the terminal itself. The position data is represented by the measurement ID associated with the latitude, longitude, and measurement time. The generated position data is stored in a storage unit in the terminal.

  When a position data transmission request arrives from the behavior prediction device SV, the GPS data transmission control unit reads the position data stored in the storage unit and transmits it to the requesting behavior prediction device SV. Note that the position data may be read from the storage unit and transmitted to the behavior prediction device SV periodically or when the latitude and longitude of the terminal changes by a certain amount or more.

  The schedule management function generates and stores scheduler data in accordance with a user input operation. When a scheduler data transmission request arrives from the behavior prediction device SV, the stored scheduler data is read out and transmitted to the requesting behavior prediction device SV. Note that the scheduler data may be read from the storage unit and transmitted to the behavior prediction apparatus SV periodically or whenever the registered content of the event is updated.

By the way, the behavior prediction apparatus SV is composed of, for example, a service server operated by a communication carrier or a service carrier, and is configured as follows. FIG. 2 is a block diagram showing the functional configuration.
That is, the behavior prediction apparatus SV includes a transmission / reception unit 1, a control unit 2, and a storage unit 3. The transmission / reception unit 1 transmits / receives information to / from the communication network NW under the control of the control unit 2.

  The storage unit 3 uses a randomly accessible non-volatile memory using an HDD (Hard Disc Drive) or an SSD (Solid State Drive). As a storage unit necessary to realize the present invention, stay behavior data A storage unit 31, a state transition sequence storage unit 32, a scheduler data storage unit 33, an event sequence data storage unit 34, and a learning data storage unit 35 are provided.

  The stay behavior data storage unit 31 is used to store a set of user stay behavior data generated by the stay location extraction processing unit 22 described later based on the position data acquired from the mobile terminals MS1 to MSn. . The staying behavior data of one user is expressed as identification information (staying place ID) indicating a staying place and information indicating the staying start time and staying end time associated with the staying action ID.

  The state transition sequence storage unit 32 associates the state transition sequence generated based on the stay behavior data stored in the stay behavior data storage unit 31 with the user ID by the state transition sequence creation processing unit 23 described later. Used to remember. The state transition sequence of one user is represented as a state transition time ID associated with a state and time.

  The scheduler data storage unit 33 is used to store the scheduler data acquired from the mobile terminals MS1 to MSn by the scheduler data acquisition control unit 24 described later in association with the user ID. The scheduler data of one user is represented as the scheduler data ID associated with the information on the character string representing the action schedule (event) and the information representing the start time and the end time.

  The event series data storage unit 34 is used to store event series data generated by the scheduler data conversion unit 25 described later based on the scheduler data stored in the scheduler data storage unit 33. One event series data is represented as an event ID associated with an event content keyword and an event start time and end time.

  The learning data storage unit 35 is used to store a probability distribution parameter of a DBN (Dynamic Bayesian Networks) model for each user, which is obtained by a learning calculation by a parameter learning processing unit 26 described later.

  The control unit 2 includes a central processing unit (CPU) as a core. As a control function necessary for carrying out the present invention, a position data acquisition control unit 21, a stay location extraction processing unit 22, A state transition sequence creation processing unit 23, a scheduler data acquisition control unit 24, a scheduler data conversion unit 25, a parameter learning processing unit 26, a behavior prediction processing unit 27, and a prediction result output control unit 28. . All of these control functions are realized by causing the CPU to execute an application program.

  The position data collection processing unit 21 controls the transmission / reception unit 1 to periodically transmit position data transmission requests to the respective mobile terminals MS1 to MSn at different timings. A process of receiving position data transmitted from MSn is executed.

  The stay location extraction processing unit 22 converts the location data acquired by the location data collection processing unit 21 into a set of stay behaviors in which the location where the user stayed and the time spent at the location are paired by clustering processing. To do. A different stay action ID is assigned to the stay action set each time the stay place changes, and the stay action data storage unit 31 stores the set of stay action data to which the stay action ID is assigned. Process.

  The state transition sequence creation processing unit 23 reads a set of stay behavior data from the stay behavior data storage unit 31, and performs a process of generating a state transition sequence based on the read set of stay behavior data. The state transition series is defined as a sample of staying behavior data at fixed time intervals and a list of user states at the sampled times. At this time, there are two types of states: staying in a certain staying place or transitioning from one staying place to another staying place. The state transition series is stored in the state transition series storage unit 32.

The scheduler data acquisition control unit 24 receives scheduler data previously created and stored by the user from the mobile terminals MS1 to MSn by the transmission / reception unit 1, and stores the received scheduler data in the scheduler data storage unit 33. I do.
The scheduler data conversion unit 25 converts the scheduler data read from the scheduler data storage unit 33 into event series data by a predetermined pre-process, and stores the converted event series data in the event series data storage unit 34. Do. The pre-processing is processing for extracting a noun by performing a morphological analysis on an arbitrary character string included in scheduler data, expressing the extracted noun as a vector, and using this as data representing schedule contents.

  The parameter learning processing unit 26 reads out the state transition series from the state transition series storage unit 32 and reads out the event series data from the event series data storage unit 34 for each user ID, and performs action prediction based on these data. The DBN model to be used is learned. The DBN model is a time series model that represents a probabilistic dependency relationship between events, and estimates parameters of a probability distribution used in the model from learning data. The parameter learning processing unit 26 stores the estimated probability distribution parameter in the learning data storage unit 35 as data used for behavior prediction.

The behavior prediction processing unit 27 has the following processing functions.
(1) The position data acquisition control unit 21 is activated to acquire data representing the current position and time of the user from the portable terminal. Then, the stay location extraction processing unit 22 and the state transition series creation processing unit 23 are activated to generate stay action data from the acquired position data, and further generate state transition series data based on the stay action data. processing.
(2) The scheduler data acquisition control unit 24 is activated to acquire user schedule data in a period until a predetermined time after the prediction target time. And the process which starts the scheduler data conversion part 25 and produces | generates event series data.
(3) The probability maximum transition search process is executed using the generated state transition sequence data and event sequence data and the probability distribution parameters stored in the learning data storage unit 35. The probabilistic maximum transition search process is a process for obtaining a state transition sequence that gives the maximum probability in the DBN model with the time zone of the future scheduler data and the start state of the event as inputs, and the state that gives this maximum probability The transition series is output as information representing the future behavior prediction result of the user.

  The prediction result output control unit 28 transmits / receives information representing the future behavior prediction result of the user obtained by the behavior prediction processing unit 27 to the server or terminal of the service provider that has a contract for using the prediction result. The process of transmitting from 1 is performed.

  Next, the operation of the behavior prediction apparatus configured as described above is divided into a learning phase in which learning processing necessary for prediction is performed and a prediction phase in which user behavior is predicted based on the learned results. To do.

(1) Learning phase In the learning phase, the following processing is performed. FIG. 3 is a flowchart showing a processing procedure and processing contents in the learning phase.
(1-1) Position Data Measurement Processing In the mobile terminals MS1 to MSn possessed by each user, position information detection and transmission processing are performed as follows in a steady state. That is, in a steady state, the mobile terminal monitors whether or not the GPS measurement timing has come to the GPS measurement timing. In this state, when, for example, 5 minutes elapses from the previous measurement timing, the GPS measurement control unit activates the GPS receiver, takes in the GPS signal received by the GPS receiver, and calculates the latitude and longitude. Then, a measurement time and a measurement ID are added to the calculated latitude and longitude as position data, and the position data is stored in a storage unit in the terminal. FIG. 8 shows an example of position data stored in the storage unit in the terminal.

  The mobile terminals MS1 to MSn monitor the reception of the position data transmission request by the position data transmission control unit while periodically performing the position data measurement process. In this state, when a transmission request is received from the behavior prediction device SV, a set of position data stored after the previous transmission is read from the storage unit in the terminal, and the set of read position data is read as the behavior prediction device. Send to SV. Thereafter, each time a transmission request is received, a set of position data stored in the storage unit in the terminal is transmitted to the behavior prediction apparatus SV.

(1-2) Position Data Acquisition Processing The control unit 2 of the behavior prediction apparatus SV first activates the position data acquisition control unit 21 in step S1, and each mobile terminal is controlled under the control of the position data acquisition control unit 21. A position data transmission request is periodically transmitted to MS1 to MSn at different timings. In response to this request, a set of position data transmitted from each of the mobile terminals MS1 to MSn is received by the transmission / reception unit 1 and stored in the position data storage unit in the storage unit.

(1-3) Extraction process of stay location When the position data for the past first period, for example, one year or several months is acquired by the position data acquisition process, the control unit 2 of the behavior prediction apparatus SV continues. In step S2, the stay location extraction processing unit 22 is activated. Then, the stay location extraction processing unit 22 reads the set of the position data, performs clustering processing on the read set of position data, and converts it into a set of stay action data. One stay behavior data is represented by a combination of a place where the user stayed, a time when the stay was started at this place, and a time when the stay was ended. In step S3, the stay location extraction processing unit 22 assigns a different stay behavior ID to each of the converted stay behavior data each time the stay location changes, and the stay behavior to which this stay behavior ID is assigned. A set of data is stored in the stay behavior storage unit 31.

As the clustering algorithm, for example, a DBSCAN algorithm that performs clustering based on the density of position data in a two-dimensional space can be applied. FIG. 9 shows a concept of processing for extracting staying behavior by clustering processing. In this example, C ₁ ,..., C ₄ represent stay places extracted. For clustering, in addition to the DBSCAN algorithm, a general clustering algorithm such as k-means or hierarchical clustering may be applied.

  When the stay place is extracted, next, the stay place extraction processing unit 22 extracts the place where the user has continuously stayed at a stay place as stay action. That is, a certain staying action is represented by an ID indicating a staying place, a time when the staying at the staying place is started, and a time when the staying is finished. An example of stay behavior data is shown in FIG. A unique stay action ID is assigned to each stay action.

  For details on the DBSCAN algorithm, see Martin Ester, Hans-Peter Kriegel, Jorg Sander, and Xiaowei Xu “A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise '', in proceedings of 2nd international conference on knowledge discovery and data mining, pp.226-231, 1996.

(1-4) Creation of State Transition Sequence Next, the control unit 2 of the behavior prediction device SV starts the state transition sequence creation processing unit 23 in step S3, and the state transition sequence creation processing unit 23 causes the stay behavior data to be generated. State transition series data is created based on a set of stay behavior data stored in the storage unit 31. The state transition series samples the stay behavior data at regular time intervals, and arranges the user state at the sampled time. The state of the user here is two types: (1) staying at a certain staying place, or (2) transitioning from one staying place to another staying place.

  FIG. 11 shows a concept when a state transition series is created from stay behavior data. When the user stays in the order of stay place IDs 3, 15, and 12, the state transition series <3, 3, 3, 3, 3 → 15, 3 → 15, 15, by sampling from the stay behavior data every 10 minutes. 15, 15, 12> are obtained. Here, 3, 15 and 12 indicate that they are staying at the corresponding stay location ID, and 3 → 15 indicates that they are in the middle of transition from the stay location 3 to 15. FIG. 12 shows an example of the state transition series data created in this way.

(1-5) Acquisition of Scheduler Data Next, the control unit 2 of the behavior prediction apparatus SV activates the scheduler data acquisition control unit 24 in step S4, and the scheduler data acquisition control unit 24 causes the mobile terminals MS1 to MSn to operate. A scheduler data transmission request is transmitted at different timings. In response to this request, scheduler data for a predetermined past period, for example, one year or several months, transmitted from each of the mobile terminals MS1 to MSn is received by the transmission / reception unit 1 and stored in the scheduler data storage unit 33.

  The scheduler data is input and stored manually by a user in a scheduler provided in the mobile terminals MS1 to MSn. As the scheduler, it is assumed that the user can enter a sentence representing the contents of the schedule and the start time and end time of the event. For example, Google Calendar (registered trademark), Microsoft (registered trademark) Office Outlook (registered trademark), and Cybozu (registered trademark) Office correspond to such a scheduler.

(1-6) Analysis of Scheduler Data Next, the control unit 2 of the behavior prediction apparatus SV starts the scheduler data conversion unit 25 in step S5, and the scheduler data conversion unit 25 extracts the scheduler data from the scheduler data storage unit 33. Read and perform analysis. In this analysis process, a string of keywords indicating event contents is created by morphologically analyzing a character string of event contents by a morphological analysis engine and extracting only nouns. Then, event data is generated by assigning an event start time and an event end time to the created keyword column of the event content. FIG. 14 shows an example of this event data.

  Next, in step S6, the scheduler data converter 25 samples the generated event data at regular time intervals to create event series data. That is, if the time is included in the time span from the start time to the end time in the event data for a sequence of times arranged at a certain time interval, the event content keyword column is associated with the time. Is a series. The generated event sequence data is stored in the event sequence data storage unit 34. FIG. 15 shows an example of this event series data. The event time of the event series data is set to be equal to the time of the state transition series data. Therefore, the user state and the event sequence are associated with the times arranged at certain time intervals.

(1-7) Parameter Learning Processing Next, the control unit 2 of the behavior prediction apparatus SV starts the parameter learning processing unit 26 in step S7, and under the control of the parameter learning processing unit 26, the event series data storage unit The event series data is read from 34, and the DBN model used for behavior prediction is learned. The DBN model is a time series model that represents a probabilistic dependency relationship between events, and estimates parameters of a probability distribution used in the model from learning data. In this embodiment, two types of models are proposed as DBN models that describe the relationship between scheduler data and position data. Hereinafter, the structure and learning procedure of each DBN model will be described in order.

(A) First DBN Model FIG. 16 shows the first DBN model. Each node in the figure represents a random variable, and the arc between nodes is a conditional probability distribution in which the distribution of the random variable represented by the node connected to the arc end point is determined by the value of the node connected to the arc start point Represents obeying. In the DBN model, time t is associated with each node. Time points are t = 0, 1, 2,..., N, where N is a constant that determines the size of the DBN model.

  FIG. 16 shows only the nodes t−2, t−1, t, and t + 1. Actually, 1, 2,..., T−3, t + 2,. Suppose you are. Each time t is associated with a corresponding time, and the difference Δt between the time associated with a certain time t and the time associated with t + 1 is always constant. The value of Δt is set to 1 minute, 10 minutes, 1 hour, etc., for example. FIG. 17 shows an example of correspondence between time t and time when Δt = 10 minutes.

ここで、Ｚ，Ｗ，Ｇは確率変数の集合であり、それぞれ・Ｚ＝｛ｚ_０，ｚ_１，…，ｚ_Ｎ｝、Ｗ＝｛Ｗ_０，Ｗ_１，…，Ｗ_Ｎ｝、Ｇ＝｛ｇ_０，ｇ_１，…，ｇ_Ｎ｝で表される。π，・Ａ，Ｂ，Ｃは確率分布を定めるパラメータであり、ＤＢＮの学習処理は学習用データからこれらのパラメータを推定する処理に相当する。各確率変数が表す対象、および式（１）の右辺中に出現する確率分布については後述する。 The first DBN model shown in FIG. 16 defines the following conditional probability distribution.

Here, Z, W, G is a set of random variables, respectively _{· Z = {z 0, z} 1, ..., z N}, W = {W 0, W 1, ..., W N}, G = {G ₀ , g ₁ ,..., G _N }. π,... A, B, and C are parameters that determine the probability distribution, and the DBN learning process corresponds to a process for estimating these parameters from the learning data. The target represented by each random variable and the probability distribution appearing in the right side of Equation (1) will be described later.

Next, the random variables z _t , W _t , and g _t that appear in the DBN model shown in FIG. 16 will be described. In the following, K types of values _{(x 1, x 2, ...} , x K) vector random variable K-dimensional taking _{x = (x 1, x 2} , ..., x K) is expressed as. x takes x _k = 1 when taking the value x _k (1 ≦ k ≦ K), and the other components have the value 0.

z _t is a random variable indicating the state of the user at time t, that is, the place where the user is staying or that the user is in the middle of transition from one place to another, and an M-dimensional vector z _t = {z _t1 , z _t2 ,..., z _tM }. M is the total number of types of states that the user can take. The total number of states that can be taken may be the total number of types of states that appear in all the state transition sequences generated in step S3 described above, or may be created by giving a set of states in advance separately from the learning data Also good. For each state, state IDs 1 to M are assigned as shown in FIG. That the user state ID of FIG. 18 in a state of m (1 ≦ m ≦ M) to a point in time t, component _{z tm} random variable _{z t} is expressed as one.

W _t is a sequence of event content keywords corresponding to time t, and W _t = {w _t1 , w _t2 ,..., W _{t | Wt |} | W _t | represents the total number of event content keywords at the time corresponding to time t. Each w _ti is a random variable representing the i-th keyword that appears at time t, and is represented by a V-dimensional vector W _ti = {w _ti1 , w _ti2 ,..., W _tiV }. V is the total number of types of keywords that can appear in the event data. The total number of keywords may be, for example, the total number of types of keywords that appear in the event series data generated in step S6, or may be created by giving a set of keywords in advance separately from the learning data.

As shown in FIG. 19, IDs 1 to V are assigned to each keyword. The i-th keyword of the event content keyword appearing at a certain time t is a keyword whose keyword ID is v (1 ≦ v ≦ V) in FIG. 19, and the component w _tiv of the random variable w _ti is 1. Represent as For example, if the event content keywords at the time corresponding to time t are “A”, “san”, and “meeting”, W _t = {w _t1 , w _t2 , w _t3 } and w _t1 , w _t2 , w _t3 is a vector in which only the component corresponding to the keyword “A” is 1, the remaining component is 0, only the component corresponding to the keyword “san” is 1, the remaining component is 0, and the keyword “meeting” Only the corresponding component is 1 and the remaining components are 0 vectors.

g _t is a random variable representing a time zone corresponding to the time t, and is represented by an R-dimensional vector g _t = {g _t1 , g _t2 ,..., g _tR }. R is the number of divisions when dividing a day. For example, when the day is divided into morning and afternoon, R = 2. For each divided time zone, a time zone ID from 1 to R is assigned as shown in FIG. When the time corresponding to a certain time t belongs to the time zone corresponding to IDr (1 ≦ r ≦ R), g _t is a vector in which only the component g _tr is 1 and the other components are 0.

として定義する。 The four types of probability distributions appearing in the above equation (1) are

Define as

を満たす。 Where π is an M-dimensional vector,

Meet.

を満たす。 A is a matrix of M rows and M columns, and for each j (1 ≦ j ≦ M)

Meet.

を満たす。 B is a matrix of M rows and V columns, and for each j (1 ≦ j ≦ M)

Meet.

を満たす。 C is a matrix of M rows and R columns, and for each j (1 ≦ j ≦ M)

Meet.

For the probability distribution, p (z ₀ | π) · is the initial distribution of states, p (z _{t + 1} | z _t , A) · is the state transition probability from time t to t + 1, and p (w _t | z _t, B) is the probability of occurrence of event content keyword string at the time · _{t, p} (g t _| respectively _z t, C) · is the probability of occurrence of the time zone at the time · t.

  Next, the procedure and contents of the parameter learning process of the first DBN model shown in FIG. 16 will be described using the flowchart of FIG. The learning process corresponds to a process for estimating the parameters A, B, and C in the above equation (1) from the event sequence data. Since the parameter π · is not used in the prediction process, no estimation is performed.

The parameter learning processing unit 26 first constructs learning data in step S71. The learning data is generated by converting the state transition sequence generated in step S3 and the event sequence data generated in step S6, and a set L = of each random variable value from t to 0 to N. {Z ₀ , z ₁ , ..., z _N , W ₀ , W ₁ , ..., W _N , g ₀ , g ₁ , ..., g _N }. An example in the case of converting learning data into a vector is shown in FIG. FIG. 21A shows the position transition series and event series data given in steps S3 and S6. With respect to this data, the time is converted with respect to the correspondence table between time t and time in FIG. 21B, the state is converted into the vector z _{t according} to FIG. Learning data FIG. 21 (f) is obtained by converting to the set W _t and converting the time zone into the time zone vector g _{t according} to FIG. 21 (d).

ここで、Ｔ_ｉｊは学習データ中でｚ_ｔｉ＝１かつｚ_{ｔ＋１．ｊ}＝１であるようなｚ_ｔ，ｚ_ｔ＋１の組が発生した回数、Ｅ_ｉｊは学習データ中で状態ｚ_ｔｉ＝１のときにキーワードｗ_ｔｋｊ＝１（ただし１≦ｋ≦｜Ｗ_ｔ｜であるようなｚ_ｔ，ｗ_ｔｋの組が発生した回数、Ｆ_ｉｊは学習データ中でｚ_ｔｉ＝１のときに時間帯ｇ_ｔｊ＝１となるようなｚ_ｔ，ｇ_ｔの組が発生した回数である。β_ｅ，β_ｇはそれぞれの確率分布のハイパーパラメータであり、学習に先立ってこれらの値を調整することでモデルの性質を調整することができる。推定されたパラメータを出力してパラメータ学習処理を終了する。 Subsequently, in step S72, the parameter learning processing unit 26 calculates the estimated values A ^, B ^, C ^ of the parameters A, B, C by the maximum likelihood estimation method based on the learning data obtained in step S71. Estimated as follows.

Here, T _ij represents z _ti = 1 and z _{t + 1. The} number of occurrences of a pair of z _t and z _{t + 1} such that _j = 1, E _ij is the keyword w _tkj = 1 (where 1 ≦ k ≦ | W _t |) when the state z _ti = 1 in the learning data The number of times such a pair of z _t and w _tk has occurred, F _ij is the number of times that a pair of z _t and g _t has occurred such that the time zone g _tj = 1 when z _ti = 1 in the learning data Β _e and β _g are hyperparameters of their respective probability distributions, and the characteristics of the model can be adjusted by adjusting these values prior to learning. The learning process ends.

である。 (B) Second DBN Model A second DBN model is shown in FIG. The probability distribution of DBN represented by this model is

It is.

What is represented by z _t , W _t , and g _t in the equation (2) is the same as the first model, and is a model in which a new node d _t is added. Let D = {d ₁ , d ₂ ,..., D _N }. π,..., A, B, C, and λ are parameters that determine the probability distribution, and these parameters are estimated from the event sequence data in the learning process as in the first DBN model.

d _t is a random variable that represents the length of time that d _t stays in state z _t and is represented by an H-dimensional vector. When d _th = 1, this means that the user stays in the current state z _t continuously for Δth time. FIG. 23 shows an example of the residence time length. Each residence time length is associated with an ID.

ここで、δ_ｉｊはｉ＝ｊのときに１を、それ以外のときに０をとる定数である。λ_ｋは状態ｚ_ｋでの滞在長さを表すパラメータであり、λはＨ次元のベクトルである。π，・Ａ，Ｂ，Ｃについては第１のＤＢＮモデルと同じである。 The right side of equation (2) is represented by the following six probability distributions.

Here, δ _ij is a constant that takes 1 when i = j and takes 0 otherwise. λ _k is a parameter representing the stay length in the state z _k , and λ is an H-dimensional vector. π, • A, B, and C are the same as those in the first DBN model.

The procedure and contents of parameter learning of the second DBN model will be described with reference to FIG.
Parameter learning processing unit 26 first _L = _{{z 0} training data in step _{S71, z 1, ..., z} N, w 0, w 1, ..., w N, g 1, g 2, ..., g N, d ₁ , d ₂ ,..., d _N }. An example of creating learning data is shown in FIG. As in the case of the learning data creation process in the first DBN model shown in FIG. 21, FIG. 24 (f) is created from FIG. 24 (a), and a new table FIG. 24 (g) is used at that time. In FIG. 24 (f), the part where the column _dt is added is different.

ここでＤ_ｋは、ｚ_{ｔ−１，ｋ}＝１かつｄ_{ｔ−１，０}＝１を満たすｄ_ｉの集合である。 Subsequently, in step S72, the parameter learning processing unit 26 performs maximum likelihood estimation of the parameter estimated values A ^, B ^, C ^, λ ^ by the following expression.

Here _{D k is} a set of _{z t-1,} k = ₁ and _{d i} satisfying _{d t-1,0} = _1.

With the above procedure, parameters can be learned from event series data by the first DBN model or the second DBN model.
Finally, in step S7, the control unit 2 stores the parameters obtained by the learning process in the learning data storage unit 35, and ends the learning process.

(2) Behavior Prediction Phase Now, when the learning process related to each user of the mobile terminals MS1 to MSn is finished, future behavior prediction of each user becomes possible.
In the behavior prediction phase, the control unit 2 of the behavior prediction device SV executes processing as follows. FIG. 5 is a flowchart showing the processing procedure and processing contents.

(2-1) Acquisition of Prediction Start State and Future Scheduler Data The control unit 2 of the behavior prediction device SV activates the behavior prediction processing unit 27, and first predicts at step S11 under the control of the behavior prediction processing unit 27. The input of information indicating the start state for starting the operation is accepted. The information indicating the start state is represented by a set of a state where the prediction is started and the time. For example, the position data acquisition control unit 21 is activated to receive the current position data from the mobile terminals MS1 to MSn, and the current position data is used as information indicating the start state. Subsequently, in step S12, the scheduler data acquisition processing unit 24 is activated, and scheduler data after the time when the user's behavior prediction is started, that is, future scheduler data, is received from the mobile terminals MS1 to MSn. In step S13, the scheduler data conversion unit 25 is activated, and morphological analysis is performed on the received scheduler data to generate event data.

(2-2) Maximum Probability Transition Search Process Next, the behavior prediction processing unit 27 executes a maximum probability transition search process in step S14. The maximum probability transition search process is a process for obtaining a series of state transitions z ₁ , z ₂ ,..., Z _N that gives the maximum probability in the DBN model with the future scheduler data, time zone, and start state as inputs. By applying a Viterbi algorithm, which is a general algorithm for obtaining a state transition sequence having the maximum probability from a hidden Markov model, it can be obtained efficiently. The process for obtaining the state transition sequence having the maximum probability differs between the first DBN model and the second DBN model, and will be described for each model.

(A) When using the first DBN model The probabilistic maximum transition search process in the first DBN model is performed as follows. FIG. 6 is a flowchart showing the processing procedure and processing contents.
That is, the action prediction processing unit 27 firstly after initializing the variable t to 0 at step S21, initializes the score for the z ₀ based on start state at step S22. The score for the state is defined as ω (z _ti ). The score in the starting state is given by ω (z _0i ) = lnp (z _0k ). That is, ω (z _0k ) = 0 for the state z _0k corresponding to the start state given as an input, and ω (z _0i ) = − ∞ for the other states.

  Next, in step S23, it is checked whether or not the value of t is equal to or less than the number of states N of the DBN model. If it is equal to or less than N, the process proceeds to step S24. In step S24, the counter i is set to 1. In step S25, it is confirmed that the counter value is equal to or less than the number M of state types. If it is M or less, the process proceeds to step S26.

として計算する。 In step S26, a score corresponding to the state z _ti is calculated. Score is

Calculate as

Next, in step S27, the previous state z _t-1j that gave the maximum value to ω (z _ti ) in step S26 is stored as ψ (z _ti ) = z _t-1j . By storing the previous state that maximizes the score in ψ (z _ti ), a sequence that maximizes the probability can be taken out later.

Subsequently, in step S28, the value of i is incremented (increased by 1), and the process proceeds to step S25. If i> M, the process proceeds to step S29, the time t is increased by 1, and the process proceeds to step S23. By repeating the above processing until time N, ω (z _ti ) and ψ (z _ti ) can be calculated for all combinations of 0 ≦ t ≦ N and 1 ≦ i ≦ M.

Next, in step S30, a state ω (z _Nk ) that takes the maximum score at time N is selected. Subsequently, z _N-1j ·· Ψ (z _Nk ) is calculated, and the calculation result is stored. Hereinafter, similarly time N-2, N-3, ..., when we seek previous state similarly Ψ by _{(z tk)} also 0, sequence _z 0 that maximizes the probability, _{z 1,} ... , Z _N are obtained.

(B) When the second DBN model is used The maximum probability transition search process in the second DBN model is performed as follows. FIG. 7 is a flowchart showing the processing procedure and processing contents. The procedure and contents of the processing are almost the same as in the case of using the first DBN model, except that the score is obtained for the set of the state z and the residence time length d in the state.

That is, the score is ω (z _ti , d _t ). In step S31, the counter t is initialized to 0, and in step S32, the score is initialized. _Assume that ω (z _0k , 0) = 0 for the initial state and ω (z _0i , d ₀ ) = − ∞ for all other combinations. Subsequently, in step S33, it is determined whether or not t is a sequence length N or less. When t is a sequence length N or less, the process proceeds to step S34.

で表される。 Next, the counters i and j are respectively initialized to 1 and 0 in step S34, and it is determined in steps S35 and S36 whether or not the values of the counters i and j exceed the state number M and the maximum time length H. To do. If the counters i and j do not exceed the number of states M and the maximum time length H, the score is calculated in step S37. The formula is

It is represented by

  Next, while increasing the value of each counter in steps S39, S40, and S41, the score and the value of Ψ are calculated for all combinations of i, j, and t. Then, after this calculation is completed, a sequence that maximizes the probability is selected in the same manner as in the case of the first DBN model in step S42, and the process ends.

(2-3) Prediction Result Output The behavior prediction processing unit 27 activates the prediction result output control unit 28 in step S15 when the probability maximum transition search process described above ends, and maximizes the obtained probability. The sequences z ₀ , z ₁ ,..., Z _N are transmitted from the transmission / reception unit 1 to a server or terminal of a service provider registered in advance.

  FIG. 25 shows an example of a prediction result, which is represented by a future time until a certain time later and a set of user states. In FIG. 25, the predicted state is represented by the staying place ID, but the coordinates (latitude / longitude) of the cluster may be output instead of the ID, and the latitude / longitude is further set to the corresponding address. The converted address information may be output after conversion.

  As described above in detail, in this embodiment, the state transition series data sampled at a constant time interval based on the past position data history acquired from the mobile terminals MS1 to MSn possessed by the user is generated and similarly. Event series data is generated based on scheduler data acquired from the mobile terminals MS1 to MSn, and the state transition series data, event series data, data representing time zones, and the residence time length in each state are used as random variables in the DBN model. Incorporating, the parameters that determine the probability distribution in the probability model are estimated. Then, the current position and future scheduler data of the user acquired from the mobile terminal and the previously estimated parameters are given to the probability model, and the state transition sequence when the probability is maximum is obtained, and the obtained state transition The series is output as information representing the user's behavior prediction result.

  Therefore, a probabilistic model representing the probabilistic coexistence between events in time series is used for predicting user behavior, and state transition sequence data generated based on the user's position data and The parameters that define the probability distribution are estimated by incorporating event series data generated based on the user's scheduler data and information representing the time zone as random variables. Therefore, by using this estimated parameter, the user's future behavior, that is, at which time and at which location, considering the history of location data and future scheduled events and time zones entered in the scheduler. It is possible to predict where to stay or from which place to move.

Furthermore, in addition to the location data history, future scheduled events entered in the scheduler and their time zones, the user's future behavior can be predicted with higher accuracy by further considering the length of stay in each state. Is possible.
The present invention is not limited to the above embodiment. For example, in the above-described embodiment, the case where the processing related to the learning phase and the processing related to the behavior prediction phase using the learning result are realized in one behavior prediction device including a server has been described as an example. However, the present invention is not limited to this, and a modeling device that performs processing related to the learning phase and a prediction device that performs processing related to the behavior prediction phase may be arranged separately. In this case, the prediction device acquires information representing the learning result from the modeling device via a communication medium such as a communication network, a signal cable, or a recording medium. Then, the user's action schedule is predicted based on the information indicating the learning result acquired from the modeling device, the position data in the past latest period separately acquired from the mobile terminal, and the scheduler data in the near future period. Perform the process.

  Moreover, in the said embodiment, the position data in the past fixed period was memorize | stored in the memory in a portable terminal, and the action prediction apparatus was made to acquire the position data for the said fixed period memorize | stored from the portable terminal. However, the present invention is not limited thereto, and each time position data is obtained, the mobile terminal may transmit this position data to the behavior prediction apparatus, and the behavior prediction apparatus may accumulate the position data.

  Furthermore, in the above-described embodiment, the case where the processing function related to the learning phase and the processing function related to the behavior prediction phase using the learning result are provided in the behavior prediction device including the server has been described as an example. You may make it provide at least one of these in a portable terminal.

  Further, the position data may be measured using the position data of the portable terminal provided by the base station of the mobile communication network in addition to the measurement using the GPS. In addition, the type and configuration of the behavior prediction device, the processing procedure and the processing content in the learning phase and the behavior prediction phase can be variously modified and implemented without departing from the gist of the present invention.

  In short, the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. Further, various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, you may combine suitably the component covering different embodiment.

  MS1 to MSn ... mobile terminal, SV ... behavior prediction device, NW ... communication network, 1 ... transmission / reception unit, 2 ... control unit, 3 ... storage unit, 21 ... location data acquisition control unit, 22 ... stay location extraction processing unit, 23 ... state transition series creation processing unit, 24 ... scheduler data acquisition control unit, 25 ... scheduler data conversion unit, 26 ... parameter learning processing unit, 27 ... behavior prediction processing unit, 28 ... prediction result output control unit, 31 ... stay behavior data Storage unit 32... State transition sequence storage unit 33. Scheduler data storage unit 34. Event sequence data storage unit 35.

Claims (8)

The position data in the past first period of the portable terminal possessed by the user is acquired, and based on the acquired position data, the sampling time set at a certain time interval and the staying state of the user at the time are expressed. Means for generating and storing state transition sequence data represented as a pair with information;
Scheduler data representing the scheduled event of the user in the first period by its contents, scheduled start time and scheduled end time is acquired, and based on the acquired scheduler data, the set sampling time and the time Means for generating and storing event series data represented as a pair with a keyword representing the content of a scheduled event in
Means for setting a plurality of time zones based on the sampling time included in each data for the stored state transition sequence data and event sequence data;
Define a probability model that represents the stochastic coexistence between events in time series for predicting the user's behavior, and read the stored state transition series data and event series data, respectively. State transition sequence data, event sequence data, and parameter learning processing means for incorporating the set time zone into the probability model as random variables and estimating and storing parameters that determine the probability distribution in the probability model A behavior prediction apparatus characterized by: The position data in the past first period of the portable terminal possessed by the user is acquired, and based on the acquired position data, the sampling time set at a certain time interval and the staying state of the user at the time are expressed. Means for generating and storing state transition sequence data represented as a pair with information;
Scheduler data representing the scheduled event of the user in the first period by its contents, scheduled start time and scheduled end time is acquired, and based on the acquired scheduler data, the set sampling time and the time Means for generating and storing event series data represented as a pair with a keyword representing the content of a scheduled event in
Means for setting a plurality of time zones based on the sampling time included in each data for the stored state transition sequence data and event sequence data;
Means for obtaining a length of time during which the user stays in each state included in the stored state transition sequence data;
Define a probability model that represents the stochastic coexistence between events in time series for predicting the user's behavior, and read the stored state transition series data and event series data, respectively. State transition sequence data, event sequence data, the set time zone, and the determined residence time length are incorporated into the probability model as random variables, respectively, and parameters for determining a probability distribution in the probability model are estimated. An action prediction apparatus comprising: a parameter learning processing means for storing.   The behavior prediction apparatus according to claim 1, wherein the parameter learning process estimates a parameter defining a probability distribution in the probability model using a maximum likelihood estimation method. Means for acquiring current position data of a mobile terminal possessed by the user and generating start state information represented by a set of the user's action start state and the time based on the acquired current position data; ,
Future scheduler data representing the scheduled event after the user's action start time is expressed by the content, scheduled start time and scheduled end time, and the content of the scheduled event is represented based on the acquired future scheduler data Means for generating event data represented as a combination of a keyword and its scheduled execution time zone;
For the probability model, the generated start state information, event data, and information representing the scheduled execution time zone are given as random variables, the stored parameters are given, and the probability corresponding to the state in the probability model is Means for obtaining a state transition sequence when maximizing;
The behavior prediction apparatus according to claim 1, further comprising means for outputting a state transition sequence that gives the obtained maximum probability as information representing the user's behavior prediction result. Means for acquiring current position data of a mobile terminal possessed by the user and generating start state information represented by a set of the user's action start state and the time based on the acquired current position data; ,
Future scheduler data representing the scheduled event after the user's action start time is expressed by the content, scheduled start time and scheduled end time, and the content of the scheduled event is represented based on the acquired future scheduler data Means for generating event data represented as a combination of a keyword and its scheduled execution time zone;
For the probability model, the generated start state information, event data, information indicating the scheduled execution time zone, and information indicating the staying time length are given as random variables, and the stored parameters are given. Means for obtaining a state transition sequence when the probability corresponding to the state and its residence time length is maximum in
3. The behavior prediction apparatus according to claim 2, further comprising means for outputting a state transition sequence that gives the maximum probability obtained as information representing the behavior prediction result of the user. The position data in the past first period of the portable terminal possessed by the user is acquired, and based on the acquired position data, the sampling time set at a certain time interval and the staying state of the user at the time are expressed. A process of generating and storing state transition sequence data represented as a pair with information;
Scheduler data representing the scheduled event of the user in the first period by its contents, scheduled start time and scheduled end time is acquired, and based on the acquired scheduler data, the set sampling time and the time A process of generating and storing event series data represented as a pair with a keyword representing the content of a scheduled event in
For the stored state transition series data and event series data, a process of setting a plurality of time zones based on sampling times included in each data,
Define a probability model that represents the stochastic coexistence between events in time series for predicting the user's behavior, and read the stored state transition series data and event series data, respectively. Including state transition sequence data, event sequence data, and a process of estimating and storing a parameter for determining a probability distribution in the probability model by incorporating the set time zone as a random variable in the probability model. Characteristic behavior prediction method. The position data in the past first period of the portable terminal possessed by the user is acquired, and based on the acquired position data, the sampling time set at a certain time interval and the staying state of the user at the time are expressed. A process of generating and storing state transition sequence data represented as a pair with information;
Scheduler data representing the scheduled event of the user in the first period by its contents, scheduled start time and scheduled end time is acquired, and based on the acquired scheduler data, the set sampling time and the time A process of generating and storing event series data represented as a pair with a keyword representing the content of a scheduled event in
For the stored state transition series data and event series data, a process of setting a plurality of time zones based on sampling times included in each data,
A process of obtaining a length of time during which the user stays in each state included in the stored state transition sequence data;
Define a probability model that represents the stochastic coexistence between events in time series for predicting the user's behavior, and read the stored state transition series data and event series data, respectively. State transition sequence data, event sequence data, the set time zone, and the determined residence time length are incorporated into the probability model as random variables, respectively, and parameters for determining a probability distribution in the probability model are estimated. A behavior prediction method comprising: a process of storing.   An action prediction program for causing a computer to execute each process included in the action scheduling method according to claim 6 or 7.

Images